Sensor-based intelligent identification and access method and system for drone accessories

Abstract

The present invention provides a sensor-based intelligent identification and access method and system for unmanned aerial vehicle accessories, which relates to the technical field of unmanned aerial vehicle accessories management, including: collecting three-dimensional contour information of unmanned aerial vehicle accessories and generating standard point cloud data, performing feature extraction and performing feature alignment and interaction, and generating a high-dimensional fusion feature matrix of target unmanned aerial vehicle accessories; constructing a multi-scale feature representation based on a feature pyramid network, generating a hierarchical feature descriptor, performing similarity calculation, generating a similarity scoring matrix, performing optimal matching, establishing an adaptive verification rule library, and performing accessory verification; calculating an optimal action sequence of an access path based on model information, mapping it to a motion trajectory parameter of a mechanical arm, collecting multi-point stress data, predicting an optimal clamping force, establishing an accessory life prediction model, combining historical data for prediction, and generating an accessory status assessment report and maintenance suggestions.

Description

Technical Field

[0001] The present invention relates to the technical field of drone accessories management, and in particular to a sensor-based drone accessories intelligent identification and access method and system. Background Art

[0002] With the rapid development of drone technology, drones have been widely used in military, agriculture, logistics, surveying and mapping and other fields. The normal operation of drones is inseparable from the support of various accessories, such as batteries, propellers, arms, etc. Efficient and accurate management of these accessories is crucial to ensure the safe flight and maintenance of drones. Traditional drone accessories management mainly adopts manual recording and regular inventory, which has problems such as low efficiency, high error rate and poor real-time performance. Most of the existing drone accessories management methods are aimed at a single type of accessories and lack a universal and highly integrated solution. Therefore, a solution is urgently needed to solve the problems existing in the existing technology. Summary of the invention

[0003] The embodiments of the present invention provide a sensor-based intelligent identification and access method and system for drone accessories, which can at least solve some of the problems existing in the prior art.

[0004] According to a first aspect of an embodiment of the present invention, a sensor-based intelligent identification and access method for drone accessories is provided, comprising: collecting three-dimensional contour information of drone accessories and performing Gaussian filtering denoising and voxel filtering downsampling processing to generate standard point cloud data, constructing an octree structure, performing three-dimensional feature matching through an iterative nearest point registration algorithm to obtain a spatial feature tensor, collecting spectral reflectance information of drone accessories, decomposing the spectral reflectance information using a continuous wavelet transform to obtain a multi-scale coefficient matrix, extracting key frequency band features based on a principal component analysis algorithm to generate a spectral feature vector, collecting surface texture information of drone accessories, generating enhanced texture data through an adaptive histogram and anisotropic diffusion filtering algorithm, inputting the spatial feature tensor, the spectral feature vector and the enhanced texture data into a pre-constructed multimodal feature fusion network, combining a cross-modal attention mechanism to perform feature alignment and interaction, and generating a high-dimensional fusion feature matrix of the target drone accessories; inputting the high-dimensional fusion feature matrix into a deep residual neural network enhanced by cascade attention, extracting key area features and channel features through a spatial attention module and a channel attention module, constructing a multi-scale feature representation based on a feature pyramid network, and generating a layer The hierarchical feature descriptor is used to calculate the similarity between the hierarchical feature descriptor and the preset standard feature library of UAV accessories according to the twin network structure to generate a similarity scoring matrix, and the similarity scoring matrix is ​​optimally matched through a dynamic programming algorithm to determine the model information and quality grade of the target UAV accessories, obtain the real-time weight data of the target UAV accessories, input the real-time weight data into a dynamic threshold learning model based on a long short-term memory network, and establish an adaptive verification rule library and perform accessory verification in combination with the standard weight range corresponding to the model information; based on the model information, a preset reinforcement learning storage location optimization model is activated, and a deep Q learning algorithm is used to calculate the optimal action sequence of the access path, and the optimal action sequence is mapped into the motion trajectory parameters of the robot arm, and a deep neural network mapping model of the material characteristics and clamping force of the target UAV accessories is established. Multi-point stress data is collected during the access process, and the optimal clamping force is predicted through a recursive neural network. After the access operation is completed, the evolution law of the accessory usage status is analyzed based on the graph attention network, and an accessory life prediction model is established. The prediction is combined with historical data to generate an accessory status assessment report and maintenance recommendations.

[0005] In an optional implementation, the three-dimensional contour information of the drone accessories is collected and Gaussian filtering denoising and voxel filtering downsampling are performed to generate standard point cloud data, an octree structure is constructed, three-dimensional feature matching is performed through an iterative nearest point registration algorithm to obtain a spatial feature tensor, spectral reflectance information of the drone accessories is collected, the spectral reflectance information is decomposed using a continuous wavelet transform to obtain a multi-scale coefficient matrix, key frequency band features are extracted based on a principal component analysis algorithm, and spectral feature vectors are generated, including: using a structured light scanner to scan the drone accessories to obtain three-dimensional point cloud data, and the three-dimensional point For each point in the cloud data, a spherical neighborhood centered on the current point is selected, and weight coefficients of all points in the spherical neighborhood are calculated based on a Gaussian function. The coordinates of all points in the spherical neighborhood and the corresponding weight coefficients are weighted averaged to obtain filtered point cloud data, and the filtered point cloud data is divided into a plurality of cubic grids with equal side lengths, each of which constitutes a voxel, and the coordinate mean of all points in each voxel is calculated as the center point coordinate of the current voxel, and all points in the current voxel are replaced with the center point coordinate to generate voxelized point cloud data as standard point cloud data; based on the The voxelized point cloud data is used to construct an octree index structure. For each point in the voxelized point cloud data, the nearest neighbor point is determined in the standard point cloud data. A least squares optimization objective function is constructed based on the nearest neighbor point. The least squares optimization objective function is solved to obtain a rigid body transformation matrix. The rigid body transformation matrix is ​​applied to the voxelized point cloud data. The process is repeated until a preset number of iterations or a convergence threshold is reached. The rigid body transformation matrix obtained in each iteration and the distances between different point pairs in the voxelized point cloud data are combined into a spatial feature tensor. A hyperspectral imager is used to image the surface of the drone accessory. Obtain visible light-near infrared band spectral data by line scanning, select Daubey wavelet basis function to perform three-layer wavelet decomposition on the visible light-near infrared band spectral data, select multiple wavelet coefficients with the largest energy in each layer of decomposition, form the selected wavelet coefficients into a multi-scale wavelet coefficient matrix, calculate the covariance matrix of the multi-scale wavelet coefficient matrix, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, select eigenvectors corresponding to the largest multiple eigenvalues ​​to construct a projection matrix, and project the multi-scale wavelet coefficient matrix onto the projection matrix to obtain spectral eigenvectors.

[0006] In an optional embodiment, the surface texture information of the drone accessory is collected, enhanced texture data is generated by an adaptive histogram and anisotropic diffusion filtering algorithm, the spatial feature tensor, the spectral feature vector and the enhanced texture data are input into a pre-constructed multimodal feature fusion network, and feature alignment and interaction are performed in combination with a cross-modal attention mechanism to generate a high-dimensional fusion feature matrix of the target drone accessory, including: using a high-resolution camera to perform multi-view imaging on the surface of the drone accessory, obtaining texture image data, calculating the pixel brightness mean of the texture image data, dividing the brightness value of each pixel in the texture image data by the pixel brightness mean to obtain a normalized texture image, dividing the normalized texture image into multiple overlapping local areas, calculating the brightness histogram of each local area, calculating the cumulative distribution function based on the brightness histogram, remapping the pixel brightness of each local area according to the cumulative distribution function to obtain an enhanced texture image, calculating the brightness gradient of the enhanced texture image in the horizontal and vertical directions, constructing a diffusion coefficient based on the brightness gradient, iteratively calculating the diffusion amount of each pixel in the orthogonal direction, and weighted combining the diffusion amount with the original pixel value to obtain an optimal texture image. texture data; constructing a multimodal deep neural network comprising three parallel encoder branches, each of the encoder branches comprising a multi-layer one-dimensional convolutional neural network, the convolution kernel size and step size of the one-dimensional convolutional neural network are set according to the dimension of the input feature, the pre-acquired spatial feature tensor, the spectral feature vector and the optimized texture data are respectively input into the corresponding encoder branches, local features are extracted by the multi-layer one-dimensional convolutional neural network, and the local features extracted by the three encoder branches are spliced ​​in the channel dimension to obtain a fused feature map; calculating the similarity between different spatial positions of the fused feature map to obtain an attention weight matrix, performing weighted summation of the attention weight matrix and the fused feature map to obtain a pooled feature map, dividing the pooled feature map into a plurality of attention heads, calculating a query matrix, a key matrix and a value matrix in each of the attention heads, calculating an attention score based on the dot product of the query matrix and the key matrix, performing weighted summation of the attention score and the value matrix to obtain a head output feature, splicing all the head output features and performing a linear transformation to obtain an attention feature, and inputting the attention feature into a fully connected layer for dimensionality reduction transformation to obtain a multimodal fused feature vector.

[0007] In an optional implementation, the high-dimensional fusion feature matrix is ​​input into a deep residual neural network enhanced by cascade attention, key area features and channel features are extracted through a spatial attention module and a channel attention module, a multi-scale feature representation is constructed based on a feature pyramid network, and a hierarchical feature descriptor is generated, including: inputting the high-dimensional fusion feature matrix into the first convolution layer of the deep residual neural network enhanced by cascade attention, obtaining a first intermediate feature through convolution operation and activation function processing, inputting the first intermediate feature into the second convolution layer to obtain a first residual term, adding the first residual term to the high-dimensional fusion feature matrix to obtain a first residual block output feature; based on the spatial attention module, performing global average pooling on the first residual block output feature in the spatial dimension to obtain a first channel descriptor, performing global maximum pooling on the spatial dimension to obtain a second channel descriptor, respectively inputting the first channel descriptor and the second channel descriptor into a multi-layer perceptron to obtain a first channel weight and a second channel weight, performing element-by-element addition on the first channel weight and the second channel weight and obtaining a channel attention vector through a sigmoid function; based on The channel attention module performs global average pooling on the output feature of the first residual block in the channel dimension to obtain a first spatial feature map, performs global maximum pooling on the channel dimension to obtain a second spatial feature map, inputs the first spatial feature map and the second spatial feature map into the convolution layer respectively to obtain a first spatial weight map and a second spatial weight map, adds the first spatial weight map and the second spatial weight map element by element and obtains a spatial attention map through a sigmoid function, multiplies the channel attention vector with the corresponding channel of the output feature of the first residual block to obtain a channel feature, multiplies the spatial attention map with the corresponding position of the output feature of the first residual block to obtain a key area feature; connects the channel feature with the key area feature in the channel dimension to obtain an enhanced feature, based on the feature pyramid network, uses the enhanced feature as the first size feature map, obtains the second size feature map through a convolution operation with a first step length of 2, and adjusts the number of channels of each feature map to 256 through a 1*1 convolution layer and performs two sections on the channel dimension to generate the hierarchical feature descriptor.

[0008] In an optional implementation, the hierarchical feature descriptor is similarly calculated with a preset standard feature library of drone accessories according to the twin network structure to generate a similarity scoring matrix, the similarity scoring matrix is ​​optimally matched through a dynamic programming algorithm, the model information and quality grade of the target drone accessories are determined, the real-time weight data of the target drone accessories are obtained, the real-time weight data is input into a dynamic threshold learning model based on a long short-term memory network, and the standard weight range corresponding to the model information is combined to establish an adaptive verification rule library and perform accessory verification, including: constructing a first feature with the same network structure and shared weight parameters The invention provides a method for extracting a feature vector and a second feature extraction subnetwork, wherein each feature extraction subnetwork comprises three fully connected layers, wherein the number of neurons in the three fully connected layers is 512, 256 and 128 respectively, wherein the hierarchical feature descriptor is input into the first feature extraction subnetwork to obtain a first output feature vector, and a standard feature vector in a preset standard feature library of drone accessories is input into the second feature extraction subnetwork to obtain a second output feature vector, and the Euclidean distance between the first output feature vector and the second output feature vector is calculated, and the Euclidean distance is mapped to a range of zero to one through an exponential function to obtain a feature similarity score; for the standard feature vector For each standard accessory in the library, the feature similarity score is repeatedly calculated to generate a similarity scoring matrix, and a state transfer matrix is ​​constructed based on the similarity scoring matrix. The number of rows and columns of the state transfer matrix are equal to the number of standard accessories in the standard feature library. Each element in the state transfer matrix represents the cost of transferring from the current state to the next state. The minimum cumulative cost to reach each state is calculated step by step based on the dynamic programming recursive formula, and the previous state node that achieves the minimum cumulative cost of each state is recorded during the calculation process to obtain the optimal predecessor node table in the state transfer process. Starting from the last state, the optimal predecessor node table is backtracked to obtain the optimal A state transition path, determining the model information and quality grade of the target drone accessories according to the terminal node of the optimal state transition path; obtaining the real-time weight data of the target drone accessories according to a high-precision electronic scale, inputting a historical weight sequence containing multiple time steps into a long short-term memory network, calculating the proportion of current input information entering the memory unit through the input gate of the long short-term memory network, calculating the proportion of historical memory information retained through the forgetting gate of the long short-term memory network, calculating the proportion of memory unit information output through the output gate of the long short-term memory network, and the long short-term memory network outputting the upper weight limit value and the lower weight limit value of multiple future time steps;According to the model information of the target drone accessory, a standard weight value and an allowable deviation value are selected from a preset adaptive verification rule library, and it is determined whether the real-time weight data is within a first weight interval determined by subtracting the allowable deviation value from the standard weight value and adding the allowable deviation value to the standard weight value, and whether the real-time weight data is within a second weight interval determined by the upper weight limit value and the lower weight limit value. If the real-time weight data is within both the first weight interval and the second weight interval, the target drone accessory is determined to be a true accessory, otherwise, the target drone accessory is determined to be a false accessory. ;

[0009] In an optional implementation, a preset reinforcement learning storage location optimization model is activated based on the model information, a deep Q learning algorithm is used to calculate the optimal action sequence of the access path, the optimal action sequence is mapped to the motion trajectory parameters of the robot arm, a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories is established, multi-point stress data is collected during the access process, and the optimal clamping force is predicted by a recursive neural network. After the access operation is completed, the evolution law of the accessory usage status is analyzed based on the graph attention network, and an accessory life prediction model is established. The prediction is combined with historical data to generate an accessory status assessment report and maintenance suggestions, including: activating a preset reinforcement learning storage location optimization model, a deep Q learning algorithm is used to calculate the optimal action sequence of the access path, and the optimal action sequence is mapped to the motion trajectory parameters of the robot arm, and a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories is established. The storage position optimization model is used to encode the model information, current position coordinates and target position coordinates of the drone accessories as a state vector, and the motion direction parameters and displacement step parameters of the robotic arm are encoded as an action vector. A Q-value neural network is constructed based on the state vector and the action vector. A reward function is constructed according to the distance value and access time value between the drone accessories and the obstacle. The Q-value neural network is trained based on the reward function to obtain an optimal action sequence. Based on the optimal action sequence, each action parameter in the optimal action sequence is mapped to the rotation angle parameters and displacement distance parameters of each joint of the robotic arm through the kinematic inverse solution algorithm to generate the motion trajectory parameters of the robotic arm. The drone is obtained. The material type information, surface roughness value and weight value of the accessory are input into a pre-trained clamping force prediction network, wherein the clamping force prediction network includes a multi-layer convolution layer for feature extraction and a fully connected layer for numerical regression, and outputs the optimal clamping force value of the robotic arm; in the process of the robotic arm executing the motion trajectory parameters, the stress data of the surface of the drone accessory is collected in real time, the stress data is input into a recursive neural network to obtain the compensated clamping force value, and the clamping force of the robotic arm is dynamically adjusted according to the compensated clamping force value to obtain the optimal clamping force; the vibration frequency value, stress The variable distribution data and the surface crack data are encoded as node features of a graph network, and the association relationship between the node features is encoded as the edge features of the graph network. The node features are adaptively weighted and aggregated through an attention mechanism to obtain a state representation vector; the state representation vector is input into a gated recurrent unit, and the future health status parameters of the drone accessory are predicted based on the historical state information and current input information of the gated recurrent unit, and a state assessment result of the drone accessory is generated according to the future health status parameters, and an accessory state assessment report and corresponding maintenance suggestions are generated based on the state assessment result, and the accessory state assessment report and the maintenance suggestions are stored in the electronic file of the drone accessory.

[0010] In an optional embodiment, each action parameter in the optimal action sequence is mapped to the rotation angle parameter and displacement distance parameter of each joint of the robotic arm through the kinematic inverse algorithm, and the motion trajectory parameters of the robotic arm are generated, including: obtaining the optimal action sequence output by the Q-value neural network, extracting the target position coordinates of the robotic arm end effector in the Cartesian space in the optimal action sequence, and the target position coordinates include X-axis coordinates, Y-axis coordinates and Z-axis coordinates; using a torque sensor to measure the first joint torque, the second joint torque and the third joint torque of the robotic arm, constructing a torque vector according to the first joint torque, the second joint torque and the third joint torque, establishing a positive kinematic mapping model of the robotic arm, and constructing a Jacobian matrix of the robotic arm based on the positive kinematic mapping model, wherein the Jacobian matrix of the robotic arm represents the mapping relationship between the joint angle change rate and the end position change rate; using a singular value decomposition algorithm The method solves the pseudo-inverse matrix of the Jacobian matrix of the robot arm, minimizes the product of the moment regularization coefficient and the torque vector as the optimization constraint target, collects the current position coordinates of the end effector of the robot arm, calculates the position error vector between the current position coordinates and the target position coordinates, and the position error vector includes an X-axis error, a Y-axis error and a Z-axis error. The position error vector and the torque vector are mapped through the pseudo-inverse matrix to obtain a first joint angle change, a second joint angle change and a third joint angle change; the first joint angle, the second joint angle and the third joint angle of the robot arm are updated according to the first joint angle change, the second joint angle change and the third joint angle change, and the update is repeated until the modulus of the position error vector is less than a preset position threshold and the modulus of the torque vector is less than a preset torque threshold, so as to obtain the motion trajectory parameters of the robot arm.

[0011] According to a second aspect of an embodiment of the present invention, a sensor-based intelligent identification and access system for drone accessories is provided, comprising: a first unit, which is used to collect three-dimensional contour information of drone accessories and perform Gaussian filtering denoising and voxel filtering downsampling processing to generate standard point cloud data, construct an octree structure, perform three-dimensional feature matching through an iterative nearest point registration algorithm to obtain a spatial feature tensor, collect spectral reflectance information of drone accessories, decompose the spectral reflectance information using a continuous wavelet transform to obtain a multi-scale coefficient matrix, extract key frequency band features based on a principal component analysis algorithm, generate a spectral feature vector, collect surface texture information of drone accessories, generate enhanced texture data through an adaptive histogram and anisotropic diffusion filtering algorithm, input the spatial feature tensor, the spectral feature vector and the enhanced texture data into a pre-constructed multimodal feature fusion network, combine the cross-modal attention mechanism to perform feature alignment and interaction, and generate a high-dimensional fusion feature matrix of the target drone accessories; a second unit, which is used to input the high-dimensional fusion feature matrix into a deep residual neural network enhanced by cascade attention, extract key area features and channel features through a spatial attention module and a channel attention module, and construct a multi-scale feature representation based on a feature pyramid network. , generate a hierarchical feature descriptor, calculate the similarity between the hierarchical feature descriptor and the preset standard feature library of drone accessories according to the twin network structure, generate a similarity scoring matrix, optimally match the similarity scoring matrix through a dynamic programming algorithm, determine the model information and quality grade of the target drone accessories, obtain the real-time weight data of the target drone accessories, input the real-time weight data into a dynamic threshold learning model based on a long short-term memory network, and establish an adaptive verification rule library and perform accessory verification in combination with the standard weight interval corresponding to the model information; the third unit is used to activate a preset reinforcement learning storage location optimization model based on the model information, calculate the optimal action sequence of the access path using a deep Q learning algorithm, map the optimal action sequence into the motion trajectory parameters of the robot arm, establish a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories, collect multi-point stress data during the access process, and predict the optimal clamping force through a recursive neural network. After the access operation is completed, analyze the evolution law of the accessory usage status based on the graph attention network, establish an accessory life prediction model, combine historical data for prediction, and generate an accessory status assessment report and maintenance recommendations.

[0012] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0014] In the present invention, by collecting multi-source heterogeneous data such as the three-dimensional contour, spectral reflectance and surface texture of drone accessories, and performing preprocessing and feature extraction, spatial feature tensors, spectral feature vectors and enhanced texture data are obtained, which comprehensively characterize the geometric, optical and surface properties of the accessories, providing a rich data basis for subsequent feature fusion and analysis. Through the spatial attention and channel attention modules, key area features and channel features are adaptively extracted. At the same time, a feature pyramid network is used to construct a multi-scale feature representation, and a hierarchical feature descriptor is generated, which greatly enhances the representation and generalization capabilities of accessory features. Based on the twin network structure and dynamic programming algorithm, the similarity calculation and optimal matching of the target accessories and the standard feature library are realized, and the accessories are accurately identified. The model information and quality grade of the accessories provide reliable information support for subsequent access management and status evaluation. The access path of the accessories is optimized through reinforcement learning and mapped to the motion trajectory parameters of the robotic arm, which realizes autonomous decision-making and intelligent planning of the access process, improves the access efficiency and accuracy, establishes a mapping model between material properties and clamping force, and predicts the optimal clamping force through a recursive neural network to ensure the safety and integrity of the accessories during the access process. In summary, the present invention realizes the automation, intelligence and refinement of the accessories management process, greatly improves the efficiency, accuracy and safety of accessories management, and provides a reliable technical means for the predictive maintenance of accessories, which is of great significance for ensuring the flight safety of drones and reducing operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The figure is a flow chart of a method for intelligently identifying and accessing drone accessories based on sensors according to an embodiment of the present invention;

[0016] Figure 2 The figure is a schematic diagram of the structure of a sensor-based intelligent identification and access system for drone accessories according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1 FIG. 1 is a flow chart of a method for intelligently identifying and accessing drone accessories based on sensors according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Collecting the three-dimensional contour information of the drone accessories and performing Gaussian filtering denoising and voxel filtering downsampling processing to generate standard point cloud data, constructing an octree structure, performing three-dimensional feature matching through an iterative nearest point registration algorithm to obtain a spatial feature tensor, collecting the spectral reflectance information of the drone accessories, decomposing the spectral reflectance information using a continuous wavelet transform to obtain a multi-scale coefficient matrix, extracting key frequency band features based on a principal component analysis algorithm, generating a spectral feature vector, collecting the surface texture information of the drone accessories, generating enhanced texture data through an adaptive histogram and anisotropic diffusion filtering algorithm, inputting the spatial feature tensor, the spectral feature vector and the enhanced texture data into a pre-constructed multimodal feature fusion network, combining the cross-modal attention mechanism for feature alignment and interaction, and generating a high-dimensional fusion feature matrix of the target drone accessories; the Gaussian filtering denoising is A commonly used image denoising technology, which removes high-frequency noise by smoothing the image with a Gaussian filter while retaining the main structure and features of the image. The voxel filtering downsampling refers to the process of voxelizing point cloud data in three-dimensional data processing to divide the point cloud into regular voxel units, and reducing the complexity and amount of calculation of the data by downsampling. The iterative closest point registration algorithm (ICP) is an algorithm for aligning two point clouds or geometric shapes, and continuously minimizes the distance between point clouds through an iterative process to ultimately achieve accurate registration. The multi-scale coefficient matrix refers to a matrix that uses coefficients at different scales to represent data features in multi-scale analysis, and is usually used in multi-resolution analysis techniques such as wavelet transform. The anisotropic diffusion filtering algorithm is an image denoising algorithm that effectively removes noise while retaining the image edge by controlling the diffusion in different directions in the image.

[0020] In an optional implementation, the three-dimensional contour information of the drone accessories is collected and Gaussian filtering denoising and voxel filtering downsampling are performed to generate standard point cloud data, an octree structure is constructed, three-dimensional feature matching is performed through an iterative nearest point registration algorithm to obtain a spatial feature tensor, spectral reflectance information of the drone accessories is collected, the spectral reflectance information is decomposed using a continuous wavelet transform to obtain a multi-scale coefficient matrix, key frequency band features are extracted based on a principal component analysis algorithm, and spectral feature vectors are generated, including: using a structured light scanner to scan the drone accessories to obtain three-dimensional point cloud data, and the three-dimensional point For each point in the cloud data, a spherical neighborhood centered on the current point is selected, and weight coefficients of all points in the spherical neighborhood are calculated based on a Gaussian function. The coordinates of all points in the spherical neighborhood and the corresponding weight coefficients are weighted averaged to obtain filtered point cloud data, and the filtered point cloud data is divided into a plurality of cubic grids with equal side lengths, each of which constitutes a voxel, and the coordinate mean of all points in each voxel is calculated as the center point coordinate of the current voxel, and all points in the current voxel are replaced with the center point coordinate to generate voxelized point cloud data as standard point cloud data; based on the The voxelized point cloud data is used to construct an octree index structure. For each point in the voxelized point cloud data, the nearest neighbor point is determined in the standard point cloud data. A least squares optimization objective function is constructed based on the nearest neighbor point. The least squares optimization objective function is solved to obtain a rigid body transformation matrix. The rigid body transformation matrix is ​​applied to the voxelized point cloud data. The process is repeated until a preset number of iterations or a convergence threshold is reached. The rigid body transformation matrix obtained in each iteration and the distances between different point pairs in the voxelized point cloud data are combined into a spatial feature tensor. A hyperspectral imager is used to image the surface of the drone accessory. Obtain visible light-near infrared band spectral data by line scanning, select Daubey wavelet basis function to perform three-layer wavelet decomposition on the visible light-near infrared band spectral data, select multiple wavelet coefficients with the largest energy in each layer of decomposition, form the selected wavelet coefficients into a multi-scale wavelet coefficient matrix, calculate the covariance matrix of the multi-scale wavelet coefficient matrix, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, select eigenvectors corresponding to the largest multiple eigenvalues ​​to construct a projection matrix, and project the multi-scale wavelet coefficient matrix onto the projection matrix to obtain spectral eigenvectors.

[0021] The spherical neighborhood refers to an area in space centered on a certain point and consisting of all points that are less than a certain radius from the point, and is commonly used in 3D point cloud data processing or image analysis. The filtered point cloud data refers to the process of applying a filter to point cloud data to smooth or denoise it, and is commonly used to improve the quality of point clouds or reduce unnecessary data redundancy. The rigid body transformation matrix is ​​a mathematical tool for representing rigid body transformations (including translation and rotation), and a 4x4 matrix is ​​usually used to represent position and direction transformations in three-dimensional space.

[0022] Use structured light scanners to obtain 3D point cloud data of drone accessories. For example, use a structured light scanner based on fringe projection technology to scan a drone propeller at 1mm intervals to obtain point cloud data containing 100,000 points.

[0023] Perform Gaussian filtering to denoise the acquired point cloud data. Take any point in the point cloud as an example and set the radius of the spherical neighborhood to 5mm. Calculate the distance from all points in the neighborhood to the center point, and convert the distance into a weight coefficient based on the Gaussian function. The closer the point is to the center point, the greater the weight coefficient. Multiply the three-dimensional coordinates of all points in the neighborhood by their corresponding weight coefficients, then divide by the sum of the weight coefficients to obtain the filtered coordinates of the point. Perform the same operation on all points to obtain the filtered point cloud data.

[0024] Perform voxel filtering and downsampling on the filtered point cloud data. Divide the space where the point cloud data is located into a cubic grid with a side length of 2mm. Each cubic grid represents a voxel. Calculate the average coordinates of all points in each voxel as the center point coordinates of the voxel. Replace all points in the voxel with the center point coordinates. Assume that after voxel filtering, the point cloud data is reduced to 20,000 points. These 20,000 points constitute the standard point cloud data.

[0025] An octree index structure is constructed based on standard point cloud data. The smallest cube containing all points is taken as the root node, and it is recursively divided into eight sub-cubes until the number of points contained in each sub-cube is less than a preset threshold (e.g., 100 points).

[0026] Perform 3D feature matching, using voxelized point cloud data as the source point cloud and standard point cloud data as the target point cloud. For each point in the source point cloud, search for its nearest neighbor in the target point cloud. Calculate the rigid body transformation matrix (including rotation and translation) between the source point cloud and the target point cloud so that the distance between the source point cloud and the target point cloud after transformation is minimized. Use the iterative closest point (ICP) algorithm for optimization. After each iteration, update the rigid body transformation matrix and calculate the distance between each point in the source point cloud and the corresponding nearest point in the target point cloud. Repeat the iteration until the preset number of iterations (e.g., 100 times) or the convergence threshold (e.g., 0.01mm) is reached. The rigid body transformation matrix and point pair distance obtained in each iteration are recorded to form a spatial feature tensor. Assume that the final spatial feature tensor dimension is 100x20000x6 (100 iterations, 20,000 points, and each iteration contains 6 parameters: 3 rotation parameters and 3 translation parameters).

[0027] Use a hyperspectral imager to obtain visible-near infrared spectral data of the surface of drone accessories. For example, with a spectral resolution of 1 nm, obtain spectral data from 400 nm to 1000 nm, a total of 601 bands.

[0028] Perform continuous wavelet transform decomposition on the acquired spectral data. Select the Daubey wavelet basis function for three-layer wavelet decomposition. In each layer of decomposition, select the first 10 wavelet coefficients with the largest energy. The selected wavelet coefficients are combined into a multi-scale wavelet coefficient matrix. Assume that the dimension of the matrix is ​​30x601 (3 layers of decomposition, 10 coefficients per layer, 601 bands).

[0029] Perform principal component analysis (PCA) on the multi-scale wavelet coefficient matrix. Calculate the covariance matrix of the matrix and perform eigenvalue decomposition on it. Select the eigenvectors corresponding to the first five largest eigenvalues ​​and construct a projection matrix. Project the multi-scale wavelet coefficient matrix onto the projection matrix to obtain a spectral eigenvector with a dimension of 30x5.

[0030] In this embodiment, the three-dimensional contour information and spectral reflectance information are integrated to more comprehensively describe the characteristics of the drone accessories and effectively improve the feature expression capability. By integrating multi-source features, different types of drone accessories can be effectively distinguished, and the accuracy of recognition and classification can be improved. Spherical neighborhood Gaussian weighted smoothing reduces the impact of noise. In summary, this embodiment realizes high-precision three-dimensional reconstruction and surface feature analysis of drone accessories, provides reliable technical support for quality inspection of drone accessories, and has important practical significance for improving the level of manufacturing quality control.

[0031] In an optional embodiment, the surface texture information of the drone accessory is collected, enhanced texture data is generated by an adaptive histogram and anisotropic diffusion filtering algorithm, the spatial feature tensor, the spectral feature vector and the enhanced texture data are input into a pre-constructed multimodal feature fusion network, and feature alignment and interaction are performed in combination with a cross-modal attention mechanism to generate a high-dimensional fusion feature matrix of the target drone accessory, including: using a high-resolution camera to perform multi-view imaging on the surface of the drone accessory, obtaining texture image data, calculating the pixel brightness mean of the texture image data, dividing the brightness value of each pixel in the texture image data by the pixel brightness mean to obtain a normalized texture image, dividing the normalized texture image into multiple overlapping local areas, calculating the brightness histogram of each local area, calculating the cumulative distribution function based on the brightness histogram, remapping the pixel brightness of each local area according to the cumulative distribution function to obtain an enhanced texture image, calculating the brightness gradient of the enhanced texture image in the horizontal and vertical directions, constructing a diffusion coefficient based on the brightness gradient, iteratively calculating the diffusion amount of each pixel in the orthogonal direction, and weighted combining the diffusion amount with the original pixel value to obtain an optimal texture image. texture data; constructing a multimodal deep neural network comprising three parallel encoder branches, each of the encoder branches comprising a multi-layer one-dimensional convolutional neural network, the convolution kernel size and step size of the one-dimensional convolutional neural network are set according to the dimension of the input feature, the pre-acquired spatial feature tensor, the spectral feature vector and the optimized texture data are respectively input into the corresponding encoder branches, local features are extracted by the multi-layer one-dimensional convolutional neural network, and the local features extracted by the three encoder branches are spliced ​​in the channel dimension to obtain a fused feature map; calculating the similarity between different spatial positions of the fused feature map to obtain an attention weight matrix, performing weighted summation of the attention weight matrix and the fused feature map to obtain a pooled feature map, dividing the pooled feature map into a plurality of attention heads, calculating a query matrix, a key matrix and a value matrix in each of the attention heads, calculating an attention score based on the dot product of the query matrix and the key matrix, performing weighted summation of the attention score and the value matrix to obtain a head output feature, splicing all the head output features and performing a linear transformation to obtain an attention feature, and inputting the attention feature into a fully connected layer for dimensionality reduction transformation to obtain a multimodal fused feature vector.

[0032] The cumulative distribution function is a function that describes the distribution of random variable values, indicating the probability that the random variable is less than or equal to a certain value. The brightness gradient refers to the rate of brightness change in an image, which is usually used for image edge detection and reflects the local change information of image brightness.

[0033] Collect the surface texture information of the drone accessories. Specifically, use a high-resolution camera (e.g., a 20-megapixel industrial camera) to image the surface of the drone accessories (e.g., propeller) from multiple perspectives (e.g., 5 perspectives: front, back, left, right, and top) to obtain texture image data. The image size obtained from each perspective is 2000x2000 pixels.

[0034] The acquired texture image data is preprocessed to calculate the average pixel brightness of the texture image data. For example, the average pixel brightness of an image is 128, and the brightness value of each pixel in the texture image data is divided by the average pixel brightness to obtain a normalized texture image. For example, if the brightness value of a pixel is 150, the normalized value is 150 / 12=1.17.

[0035] Adaptive histogram enhancement is performed on the normalized texture image. The normalized texture image is divided into multiple overlapping local areas. For example, the image is divided into overlapping areas of 64x64 pixels with an overlapping step of 32 pixels. The brightness histogram of each local area is calculated. For example, the brightness value is divided into 256 intervals. The cumulative distribution function is calculated based on the brightness histogram. The pixel brightness of each local area is remapped according to the cumulative distribution function to obtain an enhanced texture image. For example, if the cumulative distribution function value of the brightness value of a pixel in the local area is 0.8, the brightness value after remapping is 0.8*255=204.

[0036] Anisotropic diffusion filtering is performed on the enhanced texture image to calculate the brightness gradient of the enhanced texture image in the horizontal and vertical directions. For example, the gradient is calculated using the Sobel operator, and the diffusion coefficient is constructed based on the brightness gradient. For example, the gradient is mapped to the diffusion coefficient using an exponential function. The diffusion amount of each pixel in the orthogonal direction is iteratively calculated, for example, the number of iterations is set to 10 times. The diffusion amount is weightedly combined with the original pixel value, for example, the weight coefficient is set to 0.1, and the optimized texture data is obtained.

[0037] A multimodal feature fusion network is constructed, which contains three parallel encoder branches, each of which contains multiple layers of one-dimensional convolutional neural networks. The convolution kernel size and step size of the one-dimensional convolutional neural network are set according to the dimension of the input features, for example, the dimension of the spatial feature tensor is 1024, the dimension of the spectral feature vector is 512, and the dimension of the optimized texture data is 2048. One-dimensional convolution with a convolution kernel size of 3 and a step size of 1 is used for feature extraction.

[0038] Input the pre-acquired spatial feature tensor, spectral feature vector and optimized texture data into the corresponding encoder branches respectively. For example, the spatial feature tensor can be a feature extracted by a pre-trained convolutional neural network, and the spectral feature vector can be spectral data obtained by a spectrometer, and local features are extracted by a multi-layer one-dimensional convolutional neural network. The local features extracted by the three encoder branches are spliced ​​in the channel dimension to obtain a fused feature map. For example, if the output feature dimensions of the three branches are 256 respectively, the dimension of the spliced ​​fused feature map is 768.

[0039] Perform a cross-modal attention mechanism to calculate the similarity between different spatial positions of the fused feature map, for example, use the dot product to calculate the similarity and obtain the attention weight matrix. Perform a weighted summation of the attention weight matrix and the fused feature map to obtain a pooled feature map. Divide the pooled feature map into multiple attention heads, for example, into 8 attention heads. In each attention head, calculate the query matrix, key matrix, and value matrix, for example, by linear transformation. Calculate the attention score based on the dot product of the query matrix and the key matrix. Perform a weighted summation of the attention score and the value matrix to obtain the head output feature. Concatenate all the head output features and perform a linear transformation to obtain the attention feature.

[0040] The attention features are input into the fully connected layer for dimensionality reduction transformation. For example, the dimension is reduced to 512 to obtain a multimodal fusion feature vector, which is used as a high-dimensional fusion feature matrix of the target drone accessories for subsequent recognition or classification tasks.

[0041] In this embodiment, by combining spatial, spectral and enhanced texture information, the surface features of the accessories can be more comprehensively described, thereby improving the accuracy of defect detection. The adaptive histogram equalization and anisotropic diffusion filtering algorithms can effectively suppress the influence of noise and illumination changes, thereby enhancing the robustness of the algorithm. The cross-modal attention mechanism can effectively align and fuse features of different modalities, thereby improving the efficiency of feature fusion. In summary, this embodiment provides reliable technical support for the intelligent detection of drone accessories, and has important practical significance for improving the level of automation and intelligence in the manufacturing industry.

[0042] S2. Input the high-dimensional fusion feature matrix into the deep residual neural network enhanced by cascade attention, extract key area features and channel features through the spatial attention module and the channel attention module, build a multi-scale feature representation based on the feature pyramid network, generate a hierarchical feature descriptor, calculate the similarity between the hierarchical feature descriptor and the preset standard feature library of drone accessories according to the twin network structure, generate a similarity scoring matrix, optimally match the similarity scoring matrix through the dynamic programming algorithm, determine the model information and quality grade of the target drone accessories, obtain the real-time weight data of the target drone accessories, input the real-time weight data into the dynamic threshold learning model based on the long short-term memory network, and establish an adaptive verification rule library and perform accessory verification in combination with the standard weight range corresponding to the model information; the feature pyramid network is a deep learning network structure that enhances the model's recognition of targets of different scales by constructing feature maps of multiple scales. The hierarchical feature descriptor is a method of describing data features through multi-level feature extraction and representation, which is usually used in image retrieval, matching or classification tasks, and can capture multi-level information from local to global. The twin network structure is a neural network structure containing two shared parameters, which is used to compare the similarity of inputs. It is often used in image comparison and verification tasks, such as signature verification. The dynamic programming algorithm is an algorithm that reduces the amount of calculation by decomposing complex problems into simpler sub-problems and storing the results of the sub-problems. It is often used in optimization problems such as shortest path and sequence matching. The dynamic threshold learning model is a model that dynamically adjusts the threshold according to data changes. It is often used in signal processing or classification tasks to adapt to the time-varying and uncertainty of data. The adaptive verification rule base is a rule base that is automatically adjusted and optimized according to real-time data. It is often used in knowledge reasoning and automatic decision-making systems to improve the flexibility and accuracy of the system.

[0043] In an optional implementation, the high-dimensional fusion feature matrix is ​​input into a deep residual neural network enhanced by cascade attention, key area features and channel features are extracted through a spatial attention module and a channel attention module, a multi-scale feature representation is constructed based on a feature pyramid network, and a hierarchical feature descriptor is generated, including: inputting the high-dimensional fusion feature matrix into the first convolution layer of the deep residual neural network enhanced by cascade attention, obtaining a first intermediate feature through convolution operation and activation function processing, inputting the first intermediate feature into the second convolution layer to obtain a first residual term, adding the first residual term to the high-dimensional fusion feature matrix to obtain a first residual block output feature; based on the spatial attention module, performing global average pooling on the first residual block output feature in the spatial dimension to obtain a first channel descriptor, performing global maximum pooling on the spatial dimension to obtain a second channel descriptor, respectively inputting the first channel descriptor and the second channel descriptor into a multi-layer perceptron to obtain a first channel weight and a second channel weight, performing element-by-element addition on the first channel weight and the second channel weight and obtaining a channel attention vector through a sigmoid function; based on The channel attention module performs global average pooling on the output feature of the first residual block in the channel dimension to obtain a first spatial feature map, performs global maximum pooling on the channel dimension to obtain a second spatial feature map, inputs the first spatial feature map and the second spatial feature map into the convolution layer respectively to obtain a first spatial weight map and a second spatial weight map, adds the first spatial weight map and the second spatial weight map element by element and obtains a spatial attention map through a sigmoid function, multiplies the channel attention vector with the corresponding channel of the output feature of the first residual block to obtain a channel feature, multiplies the spatial attention map with the corresponding position of the output feature of the first residual block to obtain a key area feature; connects the channel feature with the key area feature in the channel dimension to obtain an enhanced feature, based on the feature pyramid network, uses the enhanced feature as the first size feature map, obtains the second size feature map through a convolution operation with a first step length of 2, and adjusts the number of channels of each feature map to 256 through a 1*1 convolution layer and performs two sections on the channel dimension to generate the hierarchical feature descriptor.

[0044] The channel descriptor is a descriptor that encodes the features of each channel in an image or signal, is often used in image processing and feature extraction, and can help identify multi-dimensional feature information.

[0045] There are many ways to obtain a high-dimensional fusion feature matrix, such as splicing the image's RGB features, depth features, texture features and other multi-modal features to form a high-dimensional matrix. For example, for a 224x224 pixel color image, its RGB feature dimension is 224x224x3, the depth feature dimension is 224x224x1, and the texture feature (such as HOG feature) dimension is 224x224x36. After splicing the features, a high-dimensional fusion feature matrix with a dimension of 224x224x40 can be obtained.

[0046] The high-dimensional fusion feature matrix is ​​input into the cascaded attention-enhanced deep residual neural network. The first convolution layer uses 64 convolution kernels of size 7x7 and stride 2 for convolution operation, and uses the ReLU activation function for nonlinear transformation to obtain the first intermediate feature. Assuming that the input feature map size is 224x224x40, after convolution and activation function processing, the dimension of the first intermediate feature is 112x112x64.

[0047] The first intermediate feature is input to the second convolution layer. The convolution layer also uses 64 convolution kernels of size 3x3 and stride 1 for convolution operation, and uses the ReLU activation function for nonlinear transformation to obtain the first residual term, which also has a dimension of 112x112x64. The first residual term is added to the high-dimensional fusion feature matrix at the pixel level. Since the dimensions of the two are inconsistent, the high-dimensional fusion feature matrix needs to be downsampled first to make its dimension consistent with the first residual term. For example, a 1x1 convolution layer can be used to adjust the number of channels of the high-dimensional fusion feature matrix from 40 to 64, and a convolution with a stride of 2 is used for downsampling so that its spatial dimension also becomes 112x112. The adjusted high-dimensional fusion feature matrix is ​​added to the first residual term to obtain the first residual block output feature with a dimension of 112x112x64.

[0048] The output features of the first residual block are subjected to global average pooling and global maximum pooling in the spatial dimension to obtain two channel descriptors with dimensions of 1x1x64. The two channel descriptors are input into two multi-layer perceptrons respectively. Each multi-layer perceptron contains two fully connected layers. The first fully connected layer reduces the feature dimension from 64 to 16, and the second fully connected layer restores the feature dimension from 16 to 64 and uses the ReLU activation function. The outputs of the two multi-layer perceptrons are added element by element, and then activated using the Sigmoid function to obtain the channel attention vector with a dimension of 1x1x64.

[0049] The output features of the first residual block are subjected to global average pooling and global maximum pooling in the channel dimension to obtain two spatial feature maps with dimensions of 112x112x1. These two spatial feature maps are input into two convolutional layers respectively, each of which uses a 7x7 convolution kernel and a ReLU activation function. The outputs of the two convolutional layers are added element by element, and then activated using the Sigmoid function to obtain a spatial attention map with dimensions of 112x112x1.

[0050] The channel attention vector is multiplied by the output feature of the first residual block at the corresponding channel to obtain the channel feature with a dimension of 112x112x64. The spatial attention map is multiplied by the output feature of the first residual block at the corresponding position to obtain the key area feature with a dimension of 112x112x64.

[0051] The channel features and key area features are concatenated in the channel dimension to obtain enhanced features with a dimension of 112x112x128. The enhanced features are used as the first-size feature map of the feature pyramid network. The first-size feature map is downsampled using a convolution operation with a step size of 2 to obtain a second-size feature map with a dimension of 56x56x128. For the first-size feature map and the second-size feature map, a 1x1 convolution layer is used to adjust the number of channels of each feature map to 256, and the first-size feature map and the second-size feature map with adjusted channel numbers are concatenated in the channel dimension to generate the final hierarchical feature descriptor with a dimension of 112x112x256+56x56x256.

[0052] In this embodiment, by combining the spatial attention module and the channel attention module, the network can pay more attention to important spatial areas and channel information, thereby extracting more discriminative features and improving feature expression capabilities. The feature pyramid network is used to construct a multi-scale feature representation, which can effectively fuse feature information of different scales, thereby improving the model's detection and recognition capabilities for targets of different sizes. Through the cascaded attention mechanism and multi-scale feature fusion, the performance of the model in various computer vision tasks can be effectively improved. In summary, this embodiment provides an innovative technical solution for industrial inspection, has important theoretical value and practical significance, and can be extended to a wider range of industrial inspection fields.

[0053] In an optional implementation, the hierarchical feature descriptor is similarly calculated with a preset standard feature library of drone accessories according to the twin network structure to generate a similarity scoring matrix, the similarity scoring matrix is ​​optimally matched through a dynamic programming algorithm, the model information and quality grade of the target drone accessories are determined, the real-time weight data of the target drone accessories are obtained, the real-time weight data is input into a dynamic threshold learning model based on a long short-term memory network, and the standard weight range corresponding to the model information is combined to establish an adaptive verification rule library and perform accessory verification, including: constructing a first feature with the same network structure and shared weight parameters The invention provides a method for extracting a feature vector and a second feature extraction subnetwork, wherein each feature extraction subnetwork comprises three fully connected layers, wherein the number of neurons in the three fully connected layers is 512, 256 and 128 respectively, wherein the hierarchical feature descriptor is input into the first feature extraction subnetwork to obtain a first output feature vector, and a standard feature vector in a preset standard feature library of drone accessories is input into the second feature extraction subnetwork to obtain a second output feature vector, and the Euclidean distance between the first output feature vector and the second output feature vector is calculated, and the Euclidean distance is mapped to a range of zero to one through an exponential function to obtain a feature similarity score; for the standard feature vector For each standard accessory in the library, the feature similarity score is repeatedly calculated to generate a similarity scoring matrix, and a state transfer matrix is ​​constructed based on the similarity scoring matrix. The number of rows and columns of the state transfer matrix are equal to the number of standard accessories in the standard feature library. Each element in the state transfer matrix represents the cost of transferring from the current state to the next state. The minimum cumulative cost to reach each state is calculated step by step based on the dynamic programming recursive formula, and the previous state node that achieves the minimum cumulative cost of each state is recorded during the calculation process to obtain the optimal predecessor node table in the state transfer process. Starting from the last state, the optimal predecessor node table is backtracked to obtain the optimal A state transition path, determining the model information and quality grade of the target drone accessories according to the terminal node of the optimal state transition path; obtaining the real-time weight data of the target drone accessories according to a high-precision electronic scale, inputting a historical weight sequence containing multiple time steps into a long short-term memory network, calculating the proportion of current input information entering the memory unit through the input gate of the long short-term memory network, calculating the proportion of historical memory information retained through the forgetting gate of the long short-term memory network, calculating the proportion of memory unit information output through the output gate of the long short-term memory network, and the long short-term memory network outputting the upper weight limit value and the lower weight limit value of multiple future time steps;According to the model information of the target drone accessory, a standard weight value and an allowable deviation value are selected from a preset adaptive verification rule library, and it is determined whether the real-time weight data is within a first weight interval determined by subtracting the allowable deviation value from the standard weight value and adding the allowable deviation value to the standard weight value, and whether the real-time weight data is within a second weight interval determined by the upper weight limit value and the lower weight limit value. If the real-time weight data is within both the first weight interval and the second weight interval, the target drone accessory is determined to be a true accessory, otherwise, the target drone accessory is determined to be a false accessory. ;

[0054] The neuron is the basic unit in the neural network, simulating the function of biological neurons, responsible for receiving input and generating output through activation functions, and is the core component of the deep learning model. The Euclidean distance is the straight-line distance between two points, which is often used to measure the similarity between two points or two vectors in space, especially in clustering, classification and image matching. The feature similarity score is a score used to measure the similarity between two features or samples, usually obtained by calculating the distance or similarity function, such as cosine similarity or Euclidean distance. The minimum cumulative cost refers to finding a path or decision sequence with the minimum cost in the optimization problem, which is often used in tasks such as path planning and image segmentation. The optimal predecessor node table is part of the dynamic programming algorithm, which is used to record the optimal path from the current node to the target node and help restore the optimal solution.

[0055] The target UAV accessories are imaged, and image processing techniques are used to extract multi-scale and multi-view image features to construct hierarchical feature descriptors. For example, the scale-invariant feature transform (SIFT) algorithm can be used to extract key points and descriptors of the image, and the local binary pattern (LBP) operator can be used to extract texture features. These features are combined into a high-dimensional vector as the hierarchical feature descriptor of the target UAV accessories.

[0056] The feature similarity is calculated using a pre-trained twin network, which contains two feature extraction subnetworks with the same structure and shared weights. Each subnetwork consists of three fully connected layers with 512, 256 and 128 neurons respectively. The hierarchical feature descriptor of the target drone accessory is input into the first subnetwork to obtain the first output feature vector. At the same time, the standard feature vectors in the preset standard feature library of drone accessories are input into the second subnetwork in sequence to obtain the corresponding second output feature vector. The Euclidean distance between the first output feature vector and each second output feature vector is calculated, and the Euclidean distance is mapped to the range of 0 to 1 through an exponential function to obtain the feature similarity score. Assuming that the output feature vector of the target accessory is [0.1, 0.2, 0.3] and the output feature vector of standard accessory A is [0.15, 0.25, 0.35], the Euclidean distance between them is about 0.0866. After exponential function mapping, the feature similarity score is about 0.92.

[0057] Repeat the above calculation process for each standard accessory in the standard feature library to generate a similarity score matrix. For example, assuming there are three standard accessories A, B, and C in the standard feature library, the generated similarity score matrix is ​​a matrix with 3 rows and 1 column, and each row represents the similarity score between the target accessory and the corresponding standard accessory. Assume that the similarity score matrix is ​​[[0.92], [0.85], [0.78]].

[0058] The state transfer matrix is ​​constructed based on the similarity score matrix. The number of rows and columns of the state transfer matrix are equal to the number of standard accessories in the standard feature library. Each element in the matrix represents the cost of transferring from the current state to the next state. Here, the negative of the similarity score can be used as the cost. For example, the state transfer matrix can be [[-0.92, 0, 0], [0, -0.85, 0], [0, 0, -0.78]]. The state transfer matrix is ​​optimally matched using the dynamic programming algorithm. The minimum cumulative cost to reach each state is calculated step by step through the dynamic programming recursive formula, and the previous state node that achieves the minimum cumulative cost of each state is recorded to obtain the optimal predecessor node table. Finally, starting from the last state, the optimal state transfer path is obtained by backtracking according to the optimal predecessor node table. The model information and quality level of the target drone accessories are determined according to the terminal node of the optimal state transfer path. Assuming that the terminal node of the optimal state transfer path is A, the model of the target drone accessories is determined to be A, and the quality level is the level corresponding to A.

[0059] Use a high-precision electronic scale to obtain the real-time weight data of the target drone accessories. Input the historical weight sequence containing multiple time steps into the pre-trained LSTM network model. For example, input the weight data of the past 10 time steps. The LSTM network controls the flow of information through the input gate, forget gate, and output gate, and outputs the upper and lower weight limits of multiple future time steps. Assume that the LSTM network outputs the upper weight limit value of the next 5 time steps as [10.2, 10.3, 10.4, 10.5, 10.6] and the lower weight limit value as [9.8, 9.7, 9.6, 9.5, 9.4].

[0060] According to the model information of the target drone accessories, select the standard weight value and the allowable deviation value from the preset adaptive verification rule library. Assuming that the standard weight value of model A is 10 grams and the allowable deviation value is 0.5 grams, the first weight interval is [9.5, 10.5]. Determine whether the real-time weight data is simultaneously in the first weight interval and the second weight interval determined by the weight upper limit and weight lower limit output by the long short-term memory network. Assuming that the current real-time weight is 10.1 grams, which is in the first weight interval [9.5, 10.5] and the second weight interval [9.8, 10.2], the target drone accessories are determined to be true accessories.

[0061] In this embodiment, the twin network structure is used to calculate the feature similarity, which can effectively extract the key features of the target drone accessories and accurately match them with the standard feature library, thereby improving the accuracy of accessory model recognition. The dynamic threshold learning model based on the long short-term memory network can predict the future weight change trend based on historical weight data, and establish an adaptive verification rule library, which effectively avoids the limitations of the fixed threshold verification method and improves the reliability of authenticity identification. By integrating the two links of model identification and weight verification, the automation and integration of drone accessories authenticity verification is realized, the verification process is simplified, and the verification efficiency is improved. In summary, this embodiment realizes the accurate determination of accessory models, quality grades and authenticity attributes, effectively solves the problems of inefficiency and subjectivity of traditional manual inspection, and is of great significance for ensuring the quality and safety of drone accessories and maintaining market order.

[0062] S3. Based on the model information, activate the preset reinforcement learning storage location optimization model, use the deep Q learning algorithm to calculate the optimal action sequence of the access path, map the optimal action sequence into the motion trajectory parameters of the robotic arm, establish a deep neural network mapping model of the material properties and clamping force of the target UAV accessories, collect multi-point stress data during the access process, predict the optimal clamping force through a recursive neural network, and when the access operation is completed, analyze the evolution of the accessory usage status based on the graph attention network, establish an accessory life prediction model, combine historical data for prediction, and generate an accessory status assessment report and maintenance recommendations.

[0063] The reinforcement learning storage location optimization model optimizes the allocation of storage locations through reinforcement learning methods to improve the efficiency of warehousing or logistics systems, usually by learning the optimal strategy to reduce access time and cost. The deep Q learning algorithm is a reinforcement learning algorithm that combines deep learning and Q learning. It approximates the Q value function through a deep neural network, so that the intelligent agent can learn the optimal behavior strategy in a high-dimensional state space. The optimal action sequence refers to the action sequence that can maximize the cumulative reward after learning in reinforcement learning, indicating the best decision path for the intelligent agent from the initial state to the target state. The optimal clamping force refers to the most appropriate clamping force calculated in order to achieve stable grasping in robot arm operation or robot control. The graph attention network is a graph neural network (GNN) that learns the relationship between nodes through an adaptive attention mechanism, thereby improving the representation ability of graph data and is widely used in tasks such as social network analysis and recommendation systems. The use state evolution law refers to predicting the future state of the system according to the state transition law of the system (such as a dynamic equation or a state transition model) for optimizing decision-making and planning.

[0064] In an optional implementation, a preset reinforcement learning storage location optimization model is activated based on the model information, a deep Q learning algorithm is used to calculate the optimal action sequence of the access path, the optimal action sequence is mapped to the motion trajectory parameters of the robot arm, a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories is established, multi-point stress data is collected during the access process, and the optimal clamping force is predicted by a recursive neural network. After the access operation is completed, the evolution law of the accessory usage status is analyzed based on the graph attention network, and an accessory life prediction model is established. The prediction is combined with historical data to generate an accessory status assessment report and maintenance suggestions, including: activating a preset reinforcement learning storage location optimization model, a deep Q learning algorithm is used to calculate the optimal action sequence of the access path, and the optimal action sequence is mapped to the motion trajectory parameters of the robot arm, and a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories is established. The storage position optimization model is used to encode the model information, current position coordinates and target position coordinates of the drone accessories as a state vector, and the motion direction parameters and displacement step parameters of the robotic arm are encoded as an action vector. A Q-value neural network is constructed based on the state vector and the action vector. A reward function is constructed according to the distance value and access time value between the drone accessories and the obstacle. The Q-value neural network is trained based on the reward function to obtain an optimal action sequence. Based on the optimal action sequence, each action parameter in the optimal action sequence is mapped to the rotation angle parameters and displacement distance parameters of each joint of the robotic arm through the kinematic inverse solution algorithm to generate the motion trajectory parameters of the robotic arm. The drone is obtained. The material type information, surface roughness value and weight value of the accessory are input into a pre-trained clamping force prediction network, wherein the clamping force prediction network includes a multi-layer convolution layer for feature extraction and a fully connected layer for numerical regression, and outputs the optimal clamping force value of the robotic arm; in the process of the robotic arm executing the motion trajectory parameters, the stress data of the surface of the drone accessory is collected in real time, the stress data is input into a recursive neural network to obtain the compensated clamping force value, and the clamping force of the robotic arm is dynamically adjusted according to the compensated clamping force value to obtain the optimal clamping force; the vibration frequency value, stress The variable distribution data and the surface crack data are encoded as node features of a graph network, and the association relationship between the node features is encoded as the edge features of the graph network. The node features are adaptively weighted and aggregated through an attention mechanism to obtain a state representation vector; the state representation vector is input into a gated recurrent unit, and the future health status parameters of the drone accessory are predicted based on the historical state information and current input information of the gated recurrent unit, and a state assessment result of the drone accessory is generated according to the future health status parameters, and an accessory state assessment report and corresponding maintenance suggestions are generated based on the state assessment result, and the accessory state assessment report and the maintenance suggestions are stored in the electronic file of the drone accessory.

[0065] The kinematic inverse algorithm is an algorithm used to solve the kinematic problems of a robot or a manipulator. By giving the position or posture of the end effector, the angle or position of each joint is calculated to achieve the required movement.

[0066] Encode the state of the accessories. Convert the model information of the drone accessories into a one-hot encoded vector, for example, "propeller-model A" is encoded as [1, 0, 0], and "motor-model B" is encoded as [0, 1, 0]. Represent the current position coordinates and target position coordinates of the accessories as three-dimensional vectors, for example, the current position [1.2, 2.5, 0.8], the target position [3.5, 1.0, 0.5], and concatenate the vectors to form a state vector. For example, the state vector can be [1, 0, 0, 1.2, 2.5, 0.8, 3.5, 1.0, 0.5].

[0067] Encode the robot's motion. Use a three-dimensional unit vector to represent the robot's motion direction parameters, for example, upward motion [0, 0, 1], forward motion [1, 0, 0]. Use a scalar to represent the displacement step length parameters, for example, moving 0.1 meters is represented as 0.1. Concatenate the direction vector and step length to form a motion vector, for example, [1, 0, 0, 0.1] represents moving 0.1 meters along the positive direction of the x-axis.

[0068] Construct a Q-value neural network and a reward function. The input of the Q-value neural network is the state vector and the action vector, and the output is the corresponding Q value. The network structure can adopt a multi-layer perceptron, for example, containing two hidden layers, each with 128 neurons. The design of the reward function takes into account the distance and access time between the accessory and the obstacle. The farther the distance to the obstacle, the shorter the access time, and the higher the reward value. For example, the reward function can be defined as the inverse of the distance plus the negative of the time.

[0069] The reward function is used to train the Q-value neural network. The deep Q-learning algorithm is used to continuously interact with the environment and update the parameters of the Q-value network to finally obtain the optimal action sequence. For example, after 1000 iterations of training, the optimal action sequence from the current position to the target position is [[1, 0, 0, 0.1], [0, 1, 0, 0.2], [0, 0, 1, 0.1]…].

[0070] According to the optimal action sequence, the motion trajectory parameters of the robot are calculated through the kinematic inverse solution algorithm. For example, for the action [1, 0, 0, 0.1], the robot's joint 1 needs to rotate 10 degrees, joint 2 needs to translate 5 centimeters, and so on, to generate the complete motion trajectory parameters.

[0071] Get the material property information of the accessory and predict the clamping force. Assume that the material type of the accessory is "aluminum alloy", the surface roughness is 0.5, and the weight is 0.2kg. Input this information into the pre-trained clamping force prediction network. The network contains multiple convolutional layers and fully connected layers. For example, the convolutional layer uses a 3x3 convolution kernel, and the fully connected layer contains 64 neurons. The network outputs the optimal clamping force value, for example, 10N.

[0072] During the access process, stress data of the accessory surface is collected in real time. For example, stress data is collected every 0.1 seconds. The collected stress data is input into a recursive neural network, such as a long short-term memory neural network, to obtain a compensated clamping force value. For example, the clamping force is adjusted to 12N according to the stress data.

[0073] Construct a graph attention network to analyze the usage status of accessories. Encode the vibration frequency, strain distribution, and surface crack data of the accessories as node features of the graph network. For example, encode the vibration frequency of 100Hz as

[100] . Encode the association between nodes as edge features. For example, the connection between two nodes is encoded as [1]. The node features are weighted and aggregated through the attention mechanism to obtain the state representation vector.

[0074] The state representation vector is input into a gated recurrent unit, such as a GRU network, to predict the future health status parameters of the accessory. For example, predict the health status in the next 10 days. Generate an accessory status assessment report and maintenance recommendations based on the prediction results. For example, the report indicates that the remaining life of the accessory is 30 days, and preventive maintenance is recommended. The report and recommendations are stored in the electronic archive of the accessory.

[0075] In this embodiment, the kinematic inverse algorithm is used to map the optimal action sequence to the rotation angle and displacement distance parameters of the robot arm joint, and a smooth and continuous robot arm motion trajectory is generated, which effectively avoids singular points and joint limit problems, improves the stability and controllability of the assembly process, accurately estimates the optimal clamping force value through end-to-end feature extraction and numerical regression, and realizes the refined control of the accessories. In the assembly execution stage, a recursive neural network is introduced to monitor and compensate the surface stress of the accessories in real time, and the clamping force of the robot arm is dynamically adjusted to ensure the force uniformity and integrity of the accessories during the assembly process, effectively reducing the risk of accessory damage, and encoding the state data of the accessories such as vibration, strain and crack into the node and edge features of the graph network. The adaptive aggregation of state information is realized through the attention mechanism, and the global state representation vector is obtained, which provides information-rich feature input for subsequent health status prediction. In summary, this embodiment realizes the autonomous optimization of the accessory assembly process and the online real-time analysis of state evaluation, which greatly improves the assembly efficiency, quality and safety, and provides a reliable technical means for the predictive maintenance of accessories, which is of great significance for promoting the intelligent upgrade of drone manufacturing and maintenance.

[0076] In an optional embodiment, each action parameter in the optimal action sequence is mapped to the rotation angle parameter and displacement distance parameter of each joint of the robotic arm through the kinematic inverse algorithm, and the motion trajectory parameters of the robotic arm are generated, including: obtaining the optimal action sequence output by the Q-value neural network, extracting the target position coordinates of the robotic arm end effector in the Cartesian space in the optimal action sequence, and the target position coordinates include X-axis coordinates, Y-axis coordinates and Z-axis coordinates; using a torque sensor to measure the first joint torque, the second joint torque and the third joint torque of the robotic arm, constructing a torque vector according to the first joint torque, the second joint torque and the third joint torque, establishing a positive kinematic mapping model of the robotic arm, and constructing a Jacobian matrix of the robotic arm based on the positive kinematic mapping model, wherein the Jacobian matrix of the robotic arm represents the mapping relationship between the joint angle change rate and the end position change rate; using a singular value decomposition algorithm The method solves the pseudo-inverse matrix of the Jacobian matrix of the robot arm, minimizes the product of the moment regularization coefficient and the torque vector as the optimization constraint target, collects the current position coordinates of the end effector of the robot arm, calculates the position error vector between the current position coordinates and the target position coordinates, and the position error vector includes an X-axis error, a Y-axis error and a Z-axis error. The position error vector and the torque vector are mapped through the pseudo-inverse matrix to obtain a first joint angle change, a second joint angle change and a third joint angle change; the first joint angle, the second joint angle and the third joint angle of the robot arm are updated according to the first joint angle change, the second joint angle change and the third joint angle change, and the update is repeated until the modulus of the position error vector is less than a preset position threshold and the modulus of the torque vector is less than a preset torque threshold, so as to obtain the motion trajectory parameters of the robot arm.

[0077] The robot arm Jacobian matrix is ​​a matrix that describes the relationship between the motion of the robot arm end effector and the change in the angle of each joint. It is widely used in robot control, especially in inverse kinematics and speed control. The singular value decomposition algorithm is a linear algebra method used to decompose a matrix into the product of three matrices, which contains the singular values ​​of the matrix. It is widely used in data dimensionality reduction, matrix approximation and signal processing. The pseudo-inverse matrix is ​​a generalized inverse of the matrix, which is used to solve ill-posed or over-posed linear equations, and is commonly found in least squares problems. The torque regularization coefficient is a coefficient used in mechanical system control to avoid excessive or unstable control signals. Regularization technology is used to balance the accuracy and robustness of the system.

[0078] Using the trained Q-value neural network, input the current state information to obtain the optimal action sequence under the corresponding state. Assume that the optimal action sequence output by the Q-value neural network is a series of target position coordinates of the end effector of the robot arm in Cartesian space, for example: the target position of action 1 is (10, 20, 30), the target position of action 2 is (15, 25, 35), and so on.

[0079] For each action in the action sequence, for example, the first action (10, 20, 30), its target position coordinates are extracted, that is, the X-axis coordinate is 10, the Y-axis coordinate is 20, and the Z-axis coordinate is 30.

[0080] Use the torque sensor installed on the joints of the robot arm to measure the torque values ​​of the first joint, the second joint, and the third joint respectively. Assume that the torque of the first joint is 2 Nm, the torque of the second joint is 3 Nm, and the torque of the third joint is 1 Nm. These three torque values ​​form a torque vector, that is, (2, 3, 1).

[0081] Based on the physical structure parameters of the robot arm, a forward kinematics mapping model of the robot arm is established. The model describes the corresponding relationship between the joint angles of the robot arm and the position of the end effector in Cartesian space. Using this forward kinematics model, the Jacobian matrix corresponding to the current robot arm joint angle is calculated. The Jacobian matrix represents the relationship between the rate of change of the joint angle and the rate of change of the end position.

[0082] The pseudo-inverse matrix of the Jacobian matrix is ​​calculated using the singular value decomposition algorithm. The pseudo-inverse matrix is ​​used to map the error of the end position to the change in the joint angle.

[0083] Set the moment regularization coefficient to 0.1. Multiply the moment regularization coefficient by the moment vector to get (0.2, 0.3, 0.1). Minimize this result as the optimization constraint objective.

[0084] Collect the current position coordinates of the end effector of the robot arm. Assume that the current position is (8, 18, 28). Calculate the position error between the current position and the target position (10, 20, 30) to obtain the position error vector (2, 2, 2).

[0085] The position error vector (2, 2, 2) and the torque vector (2, 3, 1) are mapped through a pseudo-inverse matrix to obtain the first joint angle change, the second joint angle change, and the third joint angle change. Assume that the results are 0.1 radians, 0.2 radians, and 0.3 radians, respectively.

[0086] According to the calculated angle change, the joint angle of the robot arm is updated. For example, assuming that the initial joint angle of the robot arm is (0.5, 0.6, 0.7) radians, the updated joint angle is (0.6, 0.8, 1.0) radians.

[0087] Repeat the above steps and continuously update the joint angles of the robot arm until the modulus of the position error vector is less than the preset position threshold (for example, 0.1) and the modulus of the torque vector is less than the preset torque threshold (for example, 0.5). Assume that after multiple iterations, the position error vector becomes (0.05, 0.03, 0.02) and the torque vector becomes (0.2, 0.3, 0.1), which meet the preset threshold conditions. At this time, record the joint angles and displacement distances of the robot arm, which are the motion trajectory parameters of the robot arm.

[0088] In this embodiment, by introducing torque feedback and Jacobian matrix pseudo-inverse control, the errors in the movement of the robot arm can be effectively compensated, the motion control accuracy of the robot arm can be improved, and the robot arm can reach the target position more accurately. Through torque regularization processing, the vibration and jitter of the robot arm in the movement process can be effectively suppressed, the motion stability of the robot arm can be enhanced, and the movement of the robot arm can be smoother. Through motion planning and kinematic inverse solution algorithms based on Q-value neural networks, a better motion trajectory can be generated, the movement time and energy consumption of the robot arm can be reduced, and work efficiency can be improved. In summary, this embodiment realizes a seamless mapping from the optimal motion sequence to the joint space trajectory, while taking into account the position error and torque constraints, and generating a dynamically feasible smooth trajectory, which greatly improves the autonomy, accuracy and stability of the robot arm movement, and provides a reliable motion control means for the refined assembly of drone accessories, which is of great significance for improving assembly efficiency and quality.

[0089] Figure 2 FIG. 1 is a schematic diagram of the structure of a sensor-based intelligent identification and access system for drone accessories according to an embodiment of the present invention. Figure 2As shown, the system includes: a first unit, which is used to collect three-dimensional contour information of drone accessories and perform Gaussian filtering denoising and voxel filtering downsampling processing to generate standard point cloud data, construct an octree structure, perform three-dimensional feature matching through an iterative nearest point registration algorithm to obtain a spatial feature tensor, collect spectral reflectance information of drone accessories, decompose the spectral reflectance information using a continuous wavelet transform to obtain a multi-scale coefficient matrix, extract key frequency band features based on a principal component analysis algorithm, generate a spectral feature vector, collect surface texture information of drone accessories, generate enhanced texture data through an adaptive histogram and anisotropic diffusion filtering algorithm, input the spatial feature tensor, the spectral feature vector and the enhanced texture data into a pre-constructed multimodal feature fusion network, combine the cross-modal attention mechanism to perform feature alignment and interaction, and generate a high-dimensional fusion feature matrix of the target drone accessories; a second unit, which is used to input the high-dimensional fusion feature matrix into a deep residual neural network enhanced by cascade attention, extract key area features and channel features through a spatial attention module and a channel attention module, construct a multi-scale feature representation based on a feature pyramid network, generate a hierarchical feature descriptor, and generate a high-dimensional fusion feature matrix according to the twin The network structure calculates the similarity between the hierarchical feature descriptor and the preset standard feature library of drone accessories to generate a similarity scoring matrix, optimally matches the similarity scoring matrix through a dynamic programming algorithm, determines the model information and quality grade of the target drone accessories, obtains the real-time weight data of the target drone accessories, inputs the real-time weight data into a dynamic threshold learning model based on a long short-term memory network, and establishes an adaptive verification rule library and performs accessory verification in combination with the standard weight interval corresponding to the model information; the third unit is used to activate a preset reinforcement learning storage location optimization model based on the model information, calculate the optimal action sequence of the access path using a deep Q learning algorithm, map the optimal action sequence into the motion trajectory parameters of the robot arm, establish a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories, collect multi-point stress data during the access process, predict the optimal clamping force through a recursive neural network, and when the access operation is completed, analyze the evolution law of the accessory usage status based on the graph attention network, establish an accessory life prediction model, combine historical data for prediction, and generate an accessory status assessment report and maintenance recommendations.

[0090] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0091] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0092] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sensor-based intelligent identification and access method for drone accessories, characterized in that: include: Collect the three-dimensional contour information of the drone accessories and perform Gaussian filtering denoising and voxel filtering downsampling processing to generate standard point cloud data, construct an octree structure, perform three-dimensional feature matching through an iterative nearest point registration algorithm to obtain a spatial feature tensor, collect the spectral reflectance information of the drone accessories, use continuous wavelet transform to decompose the spectral reflectance information to obtain a multi-scale coefficient matrix, extract key frequency band features based on the principal component analysis algorithm, generate spectral feature vectors, collect surface texture information of the drone accessories, generate enhanced texture data through an adaptive histogram and anisotropic diffusion filtering algorithm, input the spatial feature tensor, the spectral feature vector and the enhanced texture data into a pre-constructed multimodal feature fusion network, combine the cross-modal attention mechanism to perform feature alignment and interaction, and generate a high-dimensional fusion feature matrix of the target drone accessories; The high-dimensional fusion feature matrix is ​​input into a deep residual neural network enhanced by cascade attention, key area features and channel features are extracted through a spatial attention module and a channel attention module, a multi-scale feature representation is constructed based on a feature pyramid network, a hierarchical feature descriptor is generated, and similarity is calculated between the hierarchical feature descriptor and a preset standard feature library of drone accessories according to a twin network structure to generate a similarity scoring matrix, the similarity scoring matrix is ​​optimally matched through a dynamic programming algorithm, the model information and quality grade of the target drone accessories are determined, the real-time weight data of the target drone accessories are obtained, the real-time weight data is input into a dynamic threshold learning model based on a long short-term memory network, and an adaptive verification rule library is established in combination with the standard weight interval corresponding to the model information to perform accessory verification; Based on the model information, a preset reinforcement learning storage location optimization model is activated, and a deep Q-learning algorithm is used to calculate the optimal action sequence of the access path. The optimal action sequence is mapped into the motion trajectory parameters of the robotic arm, and a deep neural network mapping model of the material properties and clamping force of the target UAV accessories is established. Multi-point stress data is collected during the access process, and the optimal clamping force is predicted through a recursive neural network. After the access operation is completed, the evolution law of the accessory usage status is analyzed based on the graph attention network, and an accessory life prediction model is established. The prediction is combined with historical data to generate an accessory status assessment report and maintenance recommendations.

2. The method according to claim 1, characterized in that The three-dimensional contour information of the drone accessories is collected and Gaussian filtering denoising and voxel filtering downsampling are performed to generate standard point cloud data, and an octree structure is constructed. The three-dimensional feature matching is performed through the iterative closest point registration algorithm to obtain the spatial feature tensor. The spectral reflectance information of the drone accessories is collected, and the spectral reflectance information is decomposed by continuous wavelet transform to obtain a multi-scale coefficient matrix. The key frequency band features are extracted based on the principal component analysis algorithm, and the spectral feature vector is generated, including: A structured light scanner is used to scan the drone accessories to obtain three-dimensional point cloud data. For each point in the three-dimensional point cloud data, a spherical neighborhood centered on the current point is selected, and weight coefficients of all points in the spherical neighborhood are calculated based on a Gaussian function. The coordinates of all points in the spherical neighborhood are weighted averaged with the corresponding weight coefficients to obtain filtered point cloud data. The filtered point cloud data is divided into a plurality of cubic grids with equal sides, each of which constitutes a voxel. The coordinate mean of all points in each voxel is calculated as the center point coordinate of the current voxel, and all points in the current voxel are replaced with the center point coordinate to generate voxelized point cloud data as standard point cloud data. An octree index structure is constructed based on the voxelized point cloud data. For each point in the voxelized point cloud data, a nearest neighbor point is determined in the standard point cloud data. A least squares optimization objective function is constructed based on the nearest neighbor point. The least squares optimization objective function is solved to obtain a rigid body transformation matrix. The rigid body transformation matrix is ​​applied to the voxelized point cloud data. The process is repeated until a preset number of iterations or a convergence threshold is reached. The rigid body transformation matrix obtained in each iteration and the distances between different point pairs in the voxelized point cloud data are used to form a spatial feature tensor. A hyperspectral imager is used to scan the surface of the drone accessory to obtain visible light-near infrared band spectral data, and the Daubey wavelet basis function is selected to perform three-layer wavelet decomposition on the visible light-near infrared band spectral data. In each layer of decomposition, multiple wavelet coefficients with the largest energy are selected, and the selected wavelet coefficients are composed of a multi-scale wavelet coefficient matrix, and the covariance matrix of the multi-scale wavelet coefficient matrix is ​​calculated. The covariance matrix is ​​eigenvalue decomposed to obtain eigenvalues ​​and corresponding eigenvectors, and the eigenvectors corresponding to the largest multiple eigenvalues ​​are selected to construct a projection matrix, and the multi-scale wavelet coefficient matrix is ​​projected onto the projection matrix to obtain spectral eigenvectors.

3. The method according to claim 1, characterized in that The surface texture information of the drone accessories is collected, and enhanced texture data is generated through adaptive histogram and anisotropic diffusion filtering algorithms. The spatial feature tensor, the spectral feature vector and the enhanced texture data are input into a pre-built multimodal feature fusion network, and feature alignment and interaction are performed in combination with a cross-modal attention mechanism to generate a high-dimensional fusion feature matrix of the target drone accessories, including: A high-resolution camera is used to perform multi-view imaging on the surface of the drone accessory to obtain texture image data, a pixel brightness mean of the texture image data is calculated, the brightness value of each pixel in the texture image data is divided by the pixel brightness mean to obtain a normalized texture image, the normalized texture image is divided into a plurality of overlapping local areas, a brightness histogram of each local area is calculated, a cumulative distribution function is calculated based on the brightness histogram, the pixel brightness of each local area is remapped according to the cumulative distribution function to obtain an enhanced texture image, the brightness gradient of the enhanced texture image in the horizontal and vertical directions is calculated, a diffusion coefficient is constructed based on the brightness gradient, a diffusion amount of each pixel in an orthogonal direction is iteratively calculated, and the diffusion amount is weightedly combined with the original pixel value to obtain optimized texture data; Constructing a multimodal deep neural network comprising three parallel encoder branches, each of the encoder branches comprising a multi-layer one-dimensional convolutional neural network, the convolution kernel size and step size of the one-dimensional convolutional neural network are set according to the dimension of the input feature, the pre-acquired spatial feature tensor, the spectral feature vector and the optimized texture data are respectively input into the corresponding encoder branch, local features are extracted through the multi-layer one-dimensional convolutional neural network, and the local features extracted by the three encoder branches are spliced ​​in the channel dimension to obtain a fused feature map; Calculate the similarity between different spatial positions of the fused feature map to obtain an attention weight matrix, perform weighted summation of the attention weight matrix and the fused feature map to obtain a pooled feature map, divide the pooled feature map into multiple attention heads, calculate the query matrix, key matrix and value matrix in each attention head, calculate the attention score based on the dot product of the query matrix and the key matrix, perform weighted summation of the attention score and the value matrix to obtain a head output feature, concatenate all head output features and perform linear transformation to obtain an attention feature, input the attention feature into a fully connected layer for dimensionality reduction transformation, and obtain a multimodal fusion feature vector.

4. The method according to claim 1, characterized in that: The high-dimensional fusion feature matrix is ​​input into the cascaded attention enhanced deep residual neural network, and the key area features and channel features are extracted through the spatial attention module and the channel attention module. Constructing multi-scale feature representation based on feature pyramid network and generating hierarchical feature descriptors include: Input the high-dimensional fusion feature matrix into the first convolution layer of the cascaded attention enhanced deep residual neural network, obtain the first intermediate feature after convolution operation and activation function processing, input the first intermediate feature into the second convolution layer to obtain the first residual term, and add the first residual term to the high-dimensional fusion feature matrix to obtain the first residual block output feature; Based on the spatial attention module, the output features of the first residual block are globally average pooled in the spatial dimension to obtain a first channel descriptor, and the global maximum pooling is performed in the spatial dimension to obtain a second channel descriptor, and the first channel descriptor and the second channel descriptor are respectively input into a multi-layer perceptron to obtain a first channel weight and a second channel weight, and the first channel weight and the second channel weight are element-by-element added and subjected to a sigmoid function to obtain a channel attention vector; Based on the channel attention module, the output features of the first residual block are globally averaged pooled in the channel dimension to obtain a first spatial feature map, and the global maximum pooling is performed in the channel dimension to obtain a second spatial feature map. The first spatial feature map and the second spatial feature map are respectively input into the convolution layer to obtain a first spatial weight map and a second spatial weight map, and the first spatial weight map and the second spatial weight map are element-by-element added and subjected to a sigmoid function to obtain a spatial attention map. The channel attention vector is multiplied by the corresponding channel of the output feature of the first residual block to obtain a channel feature, and the spatial attention map is multiplied by the corresponding position of the output feature of the first residual block to obtain a key area feature; The channel feature and the key area feature are connected in the channel dimension to obtain an enhanced feature. Based on the feature pyramid network, the enhanced feature is used as a first-size feature map, and a second-size feature map is obtained through a convolution operation with a first step length of 2. For the first-size feature map and the second-size feature map, the number of channels of each feature map is adjusted to 256 through a 1*1 convolution layer and two sections are performed in the channel dimension to generate the hierarchical feature descriptor.

5. The method according to claim 1, characterized in that The hierarchical feature descriptors are similarly calculated with the preset standard feature library of drone accessories according to the twin network structure to generate a similarity scoring matrix, the similarity scoring matrix is ​​optimally matched through a dynamic programming algorithm, the model information and quality grade of the target drone accessories are determined, the real-time weight data of the target drone accessories are obtained, the real-time weight data is input into a dynamic threshold learning model based on a long short-term memory network, and the standard weight range corresponding to the model information is combined to establish an adaptive verification rule library and perform accessory verification, including: Constructing a first feature extraction subnetwork and a second feature extraction subnetwork with the same network structure and shared weight parameters, each feature extraction subnetwork comprising three fully connected layers, the number of neurons in the three fully connected layers being 512, 256 and 128 respectively, inputting the hierarchical feature descriptor into the first feature extraction subnetwork to obtain a first output feature vector, inputting a standard feature vector in a preset standard feature library of drone accessories into the second feature extraction subnetwork to obtain a second output feature vector, calculating the Euclidean distance between the first output feature vector and the second output feature vector, and mapping the Euclidean distance to a range of zero to one through an exponential function to obtain a feature similarity score; For each standard accessory in the standard feature library, the feature similarity score is repeatedly calculated to generate a similarity scoring matrix, and a state transfer matrix is ​​constructed based on the similarity scoring matrix. The number of rows and columns of the state transfer matrix are equal to the number of standard accessories in the standard feature library. Each element in the state transfer matrix represents the cost of transferring from the current state to the next state. The minimum cumulative cost to reach each state is calculated step by step based on the dynamic programming recursive formula, and the previous state node that achieves the minimum cumulative cost of each state is recorded during the calculation process to obtain the optimal predecessor node table in the state transfer process. Starting from the last state, the optimal state transfer path is obtained by backtracking according to the optimal predecessor node table, and the model information and quality level of the target drone accessory are determined according to the terminal node of the optimal state transfer path; The real-time weight data of the target drone accessories is obtained according to a high-precision electronic scale, and a historical weight sequence containing multiple time steps is input into a long short-term memory network. The proportion of current input information entering the memory unit is calculated through the input gate of the long short-term memory network, the proportion of historical memory information retained is calculated through the forget gate of the long short-term memory network, and the proportion of memory unit information output is calculated through the output gate of the long short-term memory network. The long short-term memory network outputs the upper and lower weight limits of multiple future time steps; According to the model information of the target UAV accessory, a standard weight value and an allowable deviation value are selected from a preset adaptive verification rule library to determine whether the real-time weight data is within a first weight range determined by subtracting the allowable deviation value from the standard weight value and adding the allowable deviation value to the standard weight value; and to determine whether the real-time weight data is within a second weight range determined by the upper weight limit value and the lower weight limit value. If the real-time weight data is within both the first weight range and the second weight range, the target UAV accessory is determined to be a true accessory; otherwise, the target UAV accessory is determined to be a false accessory.

6. The method according to claim 1, characterized in that Based on the model information, a preset reinforcement learning storage location optimization model is activated, and a deep Q learning algorithm is used to calculate the optimal action sequence of the access path. The optimal action sequence is mapped to the motion trajectory parameters of the robot arm, and a deep neural network mapping model of the material characteristics and clamping force of the target drone accessories is established. Multi-point stress data is collected during the access process, and the optimal clamping force is predicted through a recursive neural network. After the access operation is completed, the evolution law of the accessory usage status is analyzed based on the graph attention network, and an accessory life prediction model is established. The prediction is combined with historical data to generate an accessory status assessment report and maintenance suggestions, including: Activate a preset reinforcement learning storage position optimization model, encode the model information, current position coordinates and target position coordinates of the drone accessory as a state vector, encode the motion direction parameter and displacement step parameter of the robot arm as an action vector, construct a Q-value neural network based on the state vector and the action vector, construct a reward function according to the distance value and access time value between the drone accessory and the obstacle, and train the Q-value neural network based on the reward function to obtain an optimal action sequence; Based on the optimal action sequence, each action parameter in the optimal action sequence is mapped into a rotation angle parameter and a displacement distance parameter of each joint of the robotic arm through a kinematic inverse solution algorithm to generate a motion trajectory parameter of the robotic arm; Obtaining material type information, surface roughness value, and weight value of the drone accessory, inputting the material type information, surface roughness value, and weight value into a pre-trained clamping force prediction network, wherein the clamping force prediction network includes a multi-layer convolutional layer for feature extraction and a fully connected layer for numerical regression, and outputting an optimal clamping force value of the robot arm; In the process of the robot arm executing the motion trajectory parameters, the stress data of the surface of the drone accessory is collected in real time, the stress data is input into the recursive neural network to obtain the compensated clamping force value, and the clamping force of the robot arm is dynamically adjusted according to the compensated clamping force value to obtain the optimal clamping force; The vibration frequency value, strain distribution data and surface crack data of the drone accessories are encoded as node features of a graph network, the association relationship between the node features is encoded as edge features of the graph network, and the node features are adaptively weighted and aggregated through an attention mechanism to obtain a state representation vector; The state representation vector is input into a gated loop unit, future health status parameters of the drone accessory are predicted based on historical state information and current input information of the gated loop unit, a state assessment result of the drone accessory is generated according to the future health status parameters, a state assessment report of the accessory and corresponding maintenance suggestions are generated based on the state assessment result, and the accessory state assessment report and the maintenance suggestions are stored in an electronic file of the drone accessory.

7. The method according to claim 6, characterized in that Each action parameter in the optimal action sequence is mapped into the rotation angle parameter and displacement distance parameter of each joint of the robot arm through the kinematic inverse algorithm, and the motion trajectory parameters of the robot arm are generated, including: Obtaining an optimal action sequence output by the Q-value neural network, and extracting the target position coordinates of the end effector of the robot arm in the Cartesian space in the optimal action sequence, wherein the target position coordinates include X-axis coordinates, Y-axis coordinates, and Z-axis coordinates; A torque sensor is used to measure a first joint torque, a second joint torque and a third joint torque of the robotic arm, a torque vector is constructed according to the first joint torque, the second joint torque and the third joint torque, a positive kinematics mapping model of the robotic arm is established, and a Jacobian matrix of the robotic arm is constructed based on the positive kinematics mapping model, wherein the Jacobian matrix of the robotic arm represents a mapping relationship between a rate of change of a joint angle and a rate of change of an end position; A singular value decomposition algorithm is used to solve the pseudo-inverse matrix of the Jacobian matrix of the robotic arm, and the product of the moment regularization coefficient and the moment vector is minimized as the optimization constraint target. The current position coordinates of the end effector of the robotic arm are collected, and the position error vector between the current position coordinates and the target position coordinates is calculated, and the position error vector includes an X-axis error, a Y-axis error, and a Z-axis error. The position error vector and the moment vector are mapped through the pseudo-inverse matrix to obtain a first joint angle change, a second joint angle change, and a third joint angle change; The first joint angle, the second joint angle and the third joint angle of the robotic arm are updated according to the first joint angle change, the second joint angle change and the third joint angle change, and the updating is repeated until the modulus of the position error vector is less than a preset position threshold and the modulus of the torque vector is less than a preset torque threshold, so as to obtain the motion trajectory parameters of the robotic arm.

8. A sensor-based intelligent identification and access system for drone accessories, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the three-dimensional contour information of the drone accessories and perform Gaussian filtering denoising and voxel filtering downsampling processing to generate standard point cloud data, construct an octree structure, perform three-dimensional feature matching through an iterative nearest point registration algorithm to obtain a spatial feature tensor, collect the spectral reflectance information of the drone accessories, decompose the spectral reflectance information using a continuous wavelet transform to obtain a multi-scale coefficient matrix, extract key frequency band features based on a principal component analysis algorithm, generate a spectral feature vector, collect the surface texture information of the drone accessories, generate enhanced texture data through an adaptive histogram and anisotropic diffusion filtering algorithm, input the spatial feature tensor, the spectral feature vector and the enhanced texture data into a pre-constructed multimodal feature fusion network, combine the cross-modal attention mechanism to perform feature alignment and interaction, and generate a high-dimensional fusion feature matrix of the target drone accessories; The second unit is used to input the high-dimensional fusion feature matrix into a deep residual neural network enhanced by cascade attention, extract key area features and channel features through a spatial attention module and a channel attention module, construct a multi-scale feature representation based on a feature pyramid network, generate a hierarchical feature descriptor, calculate the similarity between the hierarchical feature descriptor and a preset standard feature library of drone accessories according to a twin network structure, generate a similarity score matrix, perform optimal matching on the similarity score matrix through a dynamic programming algorithm, determine the model information and quality grade of the target drone accessories, obtain the real-time weight data of the target drone accessories, input the real-time weight data into a dynamic threshold learning model based on a long short-term memory network, and establish an adaptive verification rule library and perform accessory verification in combination with the standard weight interval corresponding to the model information; The third unit is used to activate a preset reinforcement learning storage location optimization model based on the model information, use a deep Q learning algorithm to calculate the optimal action sequence of the access path, map the optimal action sequence into the motion trajectory parameters of the robotic arm, and establish a deep neural network mapping model of the material properties and clamping force of the target UAV accessories. Multi-point stress data is collected during the access process, and the optimal clamping force is predicted through a recursive neural network. After the access operation is completed, the evolution law of the accessory usage status is analyzed based on the graph attention network, and an accessory life prediction model is established. The prediction is combined with historical data to generate an accessory status assessment report and maintenance recommendations.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images